OpenAI Published 722 Math Papers From a Model You Cannot Use Yet

OpenAI math papers land without the model behind them

The OpenAI math papers published on October 6, 2026 are a strange kind of release: 722 new mathematical results, many with formal proofs checked by computer, all produced by a frontier model that OpenAI has not released. The company put the results in a public GitHub repository, including formalisations in Lean, a programming language that lets a machine verify each step of a proof. It also published 10 summaries of the model's reasoning, compute estimates, and statistics on the problems it attempted. What it did not publish is the model itself.

That split is worth sitting with. On one hand, mathematicians can read, check, and build on real results, which is more than a press release. On the other, nobody outside OpenAI can reproduce the work end to end, because the machine that did the proving stays behind closed doors. OpenAI says it consulted an independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study, and that it is working toward a responsible release of the model. Until then, the papers are evidence of capability, not access to it.

What the OpenAI math papers actually contain

The headline number is 722, but the more useful detail is the form. Many results arrive as Lean formalisations, which means a computer has checked the logic rather than a human referee skimming a sketch. That matters in mathematics, where a single missed case can sink a proof. Formal checking does not make a result important on its own, but it does make it trustworthy in a way a plain text answer from a chatbot never is.

OpenAI also disclosed process details that labs usually keep quiet: roughly three hours of ChatGPT Pro thinking per result on average, plus reasoning summaries for a sample of the work. That kind of transparency invites scrutiny, and scrutiny is exactly what this release needs. Independent mathematicians will now do what they always do with a bold claim: try to break it, extend it, and see which results survive contact with experts who were not in the room.

Why a math release is a capability signal

Mathematics is a harsh benchmark because it does not accept confident nonsense. A model can bluff its way through a marketing summary, but a proof either checks out or it does not. That is why labs keep using math as a proxy for reasoning quality, and why this release lands differently from another chatbot demo. If the results hold up, they suggest the unreleased model can sustain long, structured reasoning over hours, not just produce a clever first paragraph.

There is a parallel here with the current price war in small models. Claude Haiku 5.5 costs a tenth of its predecessor, and the industry is racing to make capable models cheap. OpenAI's math release points the other way: toward models that are expensive to run, slow to think, and aimed at problems where correctness is worth the compute. Both directions can be true at once. Cheap models handle volume, and heavy models handle proof.

What to watch next

Three questions decide whether this matters beyond a news cycle. First, do independent mathematicians find the results genuinely new and interesting, or mostly formalisations of things specialists already suspected? Second, does OpenAI actually release the model, or does responsible release stretch into another year of waiting? Third, does the company fund the follow-up work it promised, including workshops and conferences on AI-produced mathematics, or does the repository quietly stop updating?

The honest read: this is a real contribution wrapped in a familiar frustration. The papers are public and checkable, which deserves credit. The model is not, which limits what anyone else can learn from how it worked. GPT-6 is now everyone's ChatGPT, but the model behind these proofs is nobody's yet. Until that changes, treat the 722 papers as a strong claim with receipts attached, and wait for the mathematical community to deliver the verdict that actually counts.